[CHERI] Fix issue where tag_get was being CSE'd across free calls.

[resistor]

Fix for #38

[davidchisnall]

Seems to have introduced a lot of test failures?

[resistor]

@davidchisnall Interested in your thoughts on the code quality regressions this introduces. Modeling get_tag as a memory read guarantees that we don't CSE it across calls, but also kills CSE in a lot of other safe scenarios.

[davidchisnall]

I think we probably want to introduce two builtins, one that is and one that isn’t doing temporal checks. They’re independent so using both won’t be CSE’d.

A lot of of uses know that they can’t be deallocated and so are fine to use the non-temporal one, as long as they don’t use it before they’ve done the thing that guarantees that the object is not free. We can also maybe do some clever things later like rewrite the temporal one to non-temporal if we know that a capability has store-local permission.

[resistor]

@davidchisnall PTAL

[davidchisnall]

LGTM, @nwf ?

[nwf]

Seems sensible to me, though I cannot do much more than nod along to the LLVM changes.